Research map / Graph learning / GNN architectures

HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily

NeurIPS 2026 · Graph learning › GNN architectures · approach: GNN generalization theory · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HeterSEED, a dual-channel heterogeneous GNN that pairs a relation-aware semantic channel with a pseudo-label-guided structural heterophily channel, fused per node via a learned gate, with theoretical expressiveness and bias-reduction claims and experiments on five heterogeneous graphs.

Problem

Heterogeneous graph node classification under heterophily, where feature-similarity-driven metapath/relation aggregation propagates misleading information from heterophilic neighbors.

Approach

Decouples heterogeneous semantic aggregation from a structure-aware heterophily channel that splits metapath-induced neighbors into homophilic and heterophilic sets using pseudo-labels, weights them by metapath-count softmax, and fuses the two channels with a node-level gate plus a decoupling regularizer.

Key results

• HeterSEED reports the best Macro-F1/Micro-F1 on all five datasets in the main table, e.g. DBLP Macro-F1 94.48 vs 94.01 for SHGN (second-best row cell shown) and MAG Macro-F1 34.99 vs 33.28 for HETERO^2NET.
• On IMDB, HeterSEED reaches Macro-F1 66.71 and Micro-F1 70.60, versus 65.33 and 69.03 for HETERO^2NET.
• Ablations on DBLP and IMDB show every removed component lowers performance, with w/o SHC giving the largest drop on IMDB (65.79 Macro-F1).

Surprising findings

• Feature-similarity alignment with label homophily across metapaths is weak (R^2 = 0.205), motivating the approach.
• Several HGNN baselines (e.g., HINormer, DiffGraph, GRAIN, LINKX) run out of memory on the large MAG or RCDD datasets.

Limitations

• Theoretical expressiveness result uses a constructed symmetric graph and a specific fusion setup; the bias proposition assumes accurate pseudo-labels (delta-accurate) and linear smoothing.
• Pseudo-labels come from a linear classifier on current embeddings, so partition quality depends on initial accuracy.
• Reported gains over the strongest heterophily-aware baseline are often small (e.g., within one standard deviation on several cells).

Experimental setup

Datasets: DBLP IMDB ACM MAG RCDD

Benchmarks: DBLP IMDB ACM MAG RCDD

Metrics: Macro-F1 Micro-F1 Average Precision

Baselines: MLP GCN GAT LINKX FAGCN ACM-GCN GRAIN RGCN RGAT HAN HGT SHGN

Builds on: HAN RGCN HETERO^2NET ACM-GCN FAGCN

Closest papers by meaning

• Heterogeneous Graph Neural Network on Semantic Tree (AAAI 2025 · similarity 0.91)
• On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks (NeurIPS 2024 · similarity 0.90)
• Harnessing Language Model for Cross-Heterogeneity Graph Knowledge Transfer (AAAI 2025 · similarity 0.89)
• Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · similarity 0.89)
• HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.89)
• Contrastive Auxiliary Learning with Structure Transformation for Heterogeneous Graphs (AAAI 2025 · similarity 0.89)